Tempo analysis


For the tempo analysis I picked the song Great Day from MF DOOM, because on face value it seems like an outlier. Great Day, as many other tracks on the Madvillainy, uses samples in such a way that the beat sounds a bit wonky. I wonder if this makes it hard for the algorithm to track the beat. However, it’s still easy to tap along the beat at around 81bpm. The analyses confirm the wonkiness suspicion: both the measured and perceived tempo plots are impossible to make sense of. It’s interesting how a beat can be easy for a human to tap along, but almost impossible for a (near) state of the art algorithm to figure out.

Key analysis


Let’s analyze keys. To accomplish this, we will match the pitch classes we expect to hear in a key, with the actually played pitch classes. Let’s do this for an archetypal catchy rock tune: Smooth Sailing, from the Queens of the Stone Age. When we do this for all minor and major keys, we end up with the first plot. The brighter the tile, the more prevalent the key on the y-axis is. You will notice that at two points, near 135 and 215 seconds, many different keys are matched at once. When we look at the loudness plot these points coincide with the peaks in loudness. This makes sense: these points in the songs are mostly pitch-class-less noise.

Keys histogram


Here we see which keys occur how many times in the data set. Funilly, we some keys seem to be chiefly minor, like B and E, and some major, like G or D. This could be a limitation of the Spotify API, namely that it confuses relative minor and minor keys, like E minor and G major.

Introduction

This period I would like to investigate how song lyrics correlate with music properties such as modality, energy levels and dynamic pitch range across different genres. More specifically, the research focuses on relating the words contained by the lyrics to musical properties. I find it highly interesting to see how the mood and emotionality of a song/genre affect a sing and songwriter when writing the lyrics. I expect some obvious results, such that hip hop is generally more about the ‘hood’ than rock music, and perhaps that songs in a minor mode deal with sad topics more frequently than major mode songs.

Because the lyrical vocabulaire is extremely rich, a large, diverse dataset is of the essence. To accomplish this and to keep the corpus representative, I will put together a corpus that draws inspiration from a broad range of genres. This includes mainstream genres such as pop music, but also more obscure ones such as industrial hip hop, as unexpected yet interesting patterns may emerge. I will dissect the word usage for each genre and then compare word usages of different genres. It would be interesting to know if genres with similar lyrics have similar properties.

A list of albums per genre that make up the corpus (for now):

Pop: Midnights (Taylor Swift); WHEN WE ALL FALL ASLEEP, WHERE DO WE GO? (Billie Eilish); Dua Lipa (Dua Lipa)

Hip Hop: Madvillainy (MF DOOM, Madlib); ASTROWORLD (Travis Scott); HEROES & VILLAINS (Metro Boomin)

Alternative Rock: Elephant (The White Stripes); ..Like Clockwork (Queens of the Stone Age); Street Worms (Viagra Boys)

Industrial hip hop: The Money Store (Death Grips); OFFLINE! (JPEGMAFIA); Visions of Bodies Being Burned (clipping.)

Classical (translated to english): Wilhelmus (Marnix Van St. Aldegonde), Negende symfonie (Beethoven)

Analysis of an odd one out


In order for us to broaden our understanding of the entire dataset, it can be useful zoom in and investigate a single track. Let us investigate not any song, but one of the most iconic out of the entire dataset: Seven Nation Army. To the side, you will find a cepstrogram, with a level of detail per beat. As you can see, the timbre varies between two of the timbre properties. After listening to the song this makes sense: the electric guitar amplifier settings change between the chorus and verse. Though, the image is a bit noisy, so let’s investigate a more granular time division.

Analysis of an odd one out, a bit less specific


Now the time division is per bar. This gives a clearer image and it emphasizes our conclusion.

Song structure analysis


Let us try to delve deeper into our hypothesis that the different timbres are caused by differences in chorus and verse. To accomplish this, we will generate a self-similarity matrix for timbre and pitch. Here you see the self-similarity matrix of seven nation army for timbre. Our conclusion is made even stronger, it seems the timbre is split into a two parts, possibly.

Song structure analysis


To be certain here I present the pitch self-similarity matrix also. This also strongens our evidence that the song is split into two parts, we can now say this with full certainty. The blocky pattern can be explained by the chord changes.

Comparison of tracks within an album


Before we shall discover inter-album relations, let us commit to a single album, such that we can shape an idea what the variation within an album might look like.

Comparison of different albums


Self-similarity

Intuitively, tempo and energy are correlating factors of a song. We imagine high energy songs generally have a fast tempo. Let us investigate this thought by plotting the data for number of albums. As you can see, some albums tend to be limited in their energy range, whilst others are more or less contained.

Conclusion

What is there not to say. We live in a data driven society that harbors as many music tastes as there are colors in a van Gogh painting. But just like a van Gogh painting, you can dissect it and scrutinize the most elementary aspects, from its radiance to its perspective on the cruelties and absurdities of society. We looked at energy levels, at tempo, a variety of tracks and albums, from a point of view of strict objectiveness, one that our primordial ancestors would not even be able to fathom. One could draw an infinite number of conclusions, some might jump out more than others. What is a data driven society without conclusions?